Code Pipeline

When Kaizen needs to produce code (websites, dashboards, data visualizations, scripts), it runs through a dedicated quality pipeline.

Pipeline Flow

Executor calls create-plugin or edit-plugin
    |
    v
  Developer  -->  Writes or patches the code
    |
    v
  Syntax Check  -->  Validates code before execution (py_compile / node --check)
    |
    v
  Execute  -->  Runs the plugin with test inputs
    |
    v
  Conditional Review  -->  Inspects visual output (images, HTML)
    |
    v
  Pass / Fail  -->  Up to 3 attempts

Developer Agent

The Developer writes code with access to:

• Web search (Brave Search, web-fetch) for API documentation

• Context7 for version-specific library docs

• Code execution (run-snippet) for testing ideas

• Tool budget: Up to 8 tool calls per invocation

If the developer exhausts its tool budget without producing code, a final toolless call forces code production.

Patch Mode

On attempt 2+, the developer uses patch mode instead of rewriting the entire script:

<<<<<<< SEARCH
old code to find
=======
replacement code
>>>>>>> REPLACE

This is also used on attempt 1 for edit-plugin operations (existing working script should get surgical patches, not full rewrites).

If patch validation fails (leaked conflict markers, mismatched blocks), the pipeline falls back to a full rewrite.

Syntax Validation

Before execution, the code is validated:

• Python: py_compile checks for syntax errors

• Node.js / TypeScript: node --check validates syntax

If validation fails, the error is fed back to the developer for correction. Line-1 errors get a specific "preamble" hint (the model may have included prose before the code).

Execution

The plugin runs with test inputs derived from its input schema. If no schema is defined, it runs with an empty object {}. Output files are tracked as artifacts.

Conditional Review

Review only triggers for visual output (images, HTML, dashboards). The reviewer:

• Describes what it SEES in the output

• Uses REQUIREMENT | OBSERVED | GAP format

• Never suggests code (that's the developer's job)

Text-only output skips the review step.

Quality Controls

• MAX_PIPELINES_PER_RUN = 2: Caps total pipeline cycles per run

• Banned APIs: If execution fails due to missing APIs/imports, those are banned and injected into the developer's prompt on retry

• Version verification: pip show / npm list runs before the developer step to inject accurate version info

• Edit-plugin disk override: When editing a plugin, the orchestrator reads the actual current script from disk, preventing the executor's simplified version from replacing developer-enhanced code

Composite Skill + Plugin Pattern

For skills with code-heavy steps (charts, dashboards), the executor creates a dedicated plugin first (goes through the full pipeline), then creates the skill referencing that plugin. This ensures the code is quality-checked before the skill uses it.